OpenAI Standby Billing Cuts Idle Voice Agent Costs
GPT-Live-1.1 reached the API on November 20 with a standby mode that pauses inference on quiet calls. Idle minutes bill at two cents instead of five.
OpenAI's new standby mode prices the silences in a voice call, and that's where the savings sit.
OpenAI's GPT-Live-1.1 reached the API on November 20, and its standby mode bills idle voice sessions at two cents a minute.
The rate card tells the story. An active minute costs five cents, a standby minute costs two, and that's a 60 percent cut. An idle hour drops from 3.00 dollars to 1.20 dollars, and OpenAI bills by the second.
Standby pauses inference when a session goes quiet. A developer sets the wake policy to speech, picks an idle timeout between 5 and 60 seconds, and the server sends session.paused and session.resumed signals as the call sleeps and wakes.
A voice activity threshold setting rides along, with a default of 0.5 on a zero-to-one scale. The tuning decides how much background noise counts as a person talking again.
The discount has a hard edge. Standby works only on WebRTC sessions, so SIP phone calls and plain WebSocket connections pay full rate through every silence. Voice agents that run on phone lines don't get the cut.
The meter also keeps running elsewhere. The backend model behind the voice and any tool calls bill separately, so a standby session isn't a two-cent session once the agent starts reasoning again.
GPT-Live-1.1 has powered voice in ChatGPT Plus and Pro since October 15, and GPT-Live-1 reached the API on September 10. The endpoint is v1/live/sessions with a 128k default context window.
OpenAI says the new model handles accents, transcription and numbers better than its predecessor. It hasn't published benchmarks for those claims, so buyers can't check the improvement against a shared test. The pricing change is verifiable today. The quality change isn't.
The conversation.
Loading comments…
A little more perspective.
Reading is free. Verified members can comment and reply for $2 a month.